🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3668

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 8bd3bbc with merge base 561b50b ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

I see, cc @vkuzo should we just change all the attribute names to be public? e.g. _orig_dtype --> orig_dtype and _is_swizzled_scales --> is_swizzled_scales in NVFP4Tensor (

I see, cc @vkuzo should we just change all the attribute names to be public? e.g. _orig_dtype --> orig_dtype and _is_swizzled_scales --> is_swizzled_scales in NVFP4Tensor (

ao/torchao/prototype/mx_formats/nvfp4_tensor.py

Lines 68 to 74 in 220fedc

	tensor_attribute_names = [
	"block_size",
	"_orig_dtype",
	]
	optional_tensor_data_names = ["per_tensor_scale", "act_per_tensor_scale"]
	optional_tensor_attribute_names = [
	"_is_swizzled_scales",

)

sounds good, as long as we are clear there is no BC on this as its in prototype

yeah it's clear I think, since these are imported from prototype folder.

@liangel-02 maybe first open a PR to change these names first, then we can remove the workaround in this PR

changes looks good, please add test with checkpoints as well to make sure it works e2e (can either upload to https://huggingface.co/torchao-testing or just test with personal id)

@jerryzh168 i uploaded both checkpoints to https://huggingface.co/torchao-testing, see https://huggingface.co/torchao-testing/Qwen3-8B-NVFP4-safetensors and https://huggingface.co/torchao-testing/Qwen3-8B-MXFP8-safetensors. merging now!

@liangel-02 thanks, please update the model card of these checkpoints on repro instructions as well: https://huggingface.co/torchao-testing/Qwen3-8B-NVFP4-safetensors, example: https://huggingface.co/torchao-testing/opt-125m-Float8WeightOnlyConfig-v2-0.15.0